To set up a Ray cluster with Docker Compose (head node, 2 workers, and a Ray Serve app), you should run the head and worker nodes as separate services, then deploy your Ray Serve app connected to the head. The Ray Serve app should connect to the head node using ray.init(address='ray://ray-head:10001') and start Serve with serve.start(detached=True, http_options={"host": "0.0.0.0"}) to expose it for external access. This is a standard and recommended approach for local development and testing with Docker Compose, as shown in this example.
A minimal docker-compose.yml would define services for the head, workers, and app, each with the appropriate ports and network settings. The Ray Serve app should not run its own head node but connect to the cluster’s head. For a full example and best practices, see the detailed setup and scripts in this discussion. Would you like a step-by-step breakdown or a sample docker-compose.yml?
Sources:
Hint: Mention @RunLLM in the post for followups.